Abstract
Gait recognition has emerged as a powerful tool for unobtrusive andlong-range identity analysis, with growing relevance in surveillance andmonitoring applications. Although recent advances in deep learning andlarge-scale datasets have enabled highly accurate recognition under closed-setconditions, real-world deployment demands open-set gait enrollment, which meansdetermining whether a new gait sample corresponds to a known identity orrepresents a previously unseen individual. In this work, we introduce atransformer-based framework for open-set gait enrollment that is bothdataset-agnostic and recognition-architecture-agnostic. Our method leverages aSetTransformer to make enrollment decisions based on the embedding of a probesample and a context set drawn from the gallery, without requiringtask-specific thresholds or retraining for new environments. By decouplingenrollment from the main recognition pipeline, our model is generalized acrossdifferent datasets, gallery sizes, and identity distributions. We propose anevaluation protocol that uses existing datasets in different ratios ofidentities and walks per identity. We instantiate our method usingskeleton-based gait representations and evaluate it on two benchmark datasets(CASIA-B and PsyMo), using embeddings from three state-of-the-art recognitionmodels (GaitGraph, GaitFormer, and GaitPT). We show that our method isflexible, is able to accurately perform enrollment in different scenarios, andscales better with data compared to traditional approaches. We will make thecode and dataset scenarios publicly available.